Free tools · Sources included
Choose a model from recorded results, or estimate a cost with your own inputs. Each tool keeps measured evidence separate from a calculation.
Recorded evidence
Which kinds of memory match my model and team?
You enter: Choose a recorded model, team facts, enforcement needs and whether notes grow without review.
The map does not read outcomes or rank memory kinds. Recorded overlaps and missing values stay visible.
Calculation
Which recorded router fits my timing cap and cost basis?
You enter: Enter decisions per day, working days and a recorded median or p95 timing cap.
Recorded direct-API and CLI routes differ. This is not a quality forecast, queue model or SLA.
Recorded evidence
Which measured configuration fits my task?
You enter: Pick a recorded task, execution route and cost or speed preference.
A choice from the recorded sample, not a prediction for your work.
Calculation
What would my workload cost at API list prices?
You enter: Choose a workload or enter token counts, tasks per month and a price route.
A list-price estimate, not a subscription bill or vendor invoice.
Calculation
When could a repeated prompt pay back its cache write?
You enter: Enter a repeated prefix, new tokens, turns and sessions.
The estimate assumes the stated cache reuse. Recorded sessions remain separate.
Calculation
How much time could waiting on a model add?
You enter: Choose a recorded task; enter tasks per person per day, people and working days. Set whether someone waits and an optional hourly cost.
Scaled benchmark timing, not measured team productivity or realized savings.
Calculation
How many runs would separate two fixed pass rates?
You enter: Enter two pass rates and an optional cost per run.
An independent fixed-rate calculation, not statistical power or a prediction for your tasks.
Calculation
What could independent retries add to answer cost?
You enter: Choose two recorded hard-task configurations and a requested number of correct answers.
Fixed independent retries are an assumption; unknown costs remain unknown. This is not a vendor invoice.
Recorded evidence
How do two recorded metrics compare within one study?
You enter: Choose one study and two source-qualified metrics.
Aggregate study values, not paired raw runs or a rank. No cross-study plot.
Recorded results describe the tested tasks, routes and dates. Calculator outputs depend on your inputs and the stated prices or timing. They do not prove a bill, customer savings or future performance.
Start with our benchmark methodology, or inspect the source studies.
Crewshal runs these choices for you: a persistent AI worker with memory and rules, on your Claude and Codex subscriptions.